Projects in which our specialists have gained experience

Customer Churn Prediction

---

Customer Churn Prediction was a project for a Polish bank. Given historical data, we developed a time-series model for predicting if a client will likely churn in the following month. During development, we found out that the task given by the bank was vague and that more valuable data could be collected. We recommended future actions (such as A/B testing) and suggested developing a different, personalized activity metric that would tell the bank more about customer engagement. We also developed a customer clustering system to aid the marketing department with choosing a personalized communication strategy. The technological stack consisted of Google Cloud Platform solutions, such as VertexAI and Compute Engine.

Identification of Potential Inhibitors of the NSP-13 Protein of the SARS-CoV-2 Virus

---

Using deep learning, we successfully developed a model that aids in the identification of novel drugs capable of combating SARS-CoV-2, by discovering new enzyme inhibitors. The aim of this work was to identify potential inhibitors of the Nsp-13 protein of the SARS-CoV-2 virus using an approach based on deep learning. The input for training the neural network was a set of protein-ligand pairs. The output was a probability of a ligand-protein pair having an IC50 value < 10 μM, which creates possibility that given ligand is an inhibitor. In the end the prediction resulted in an accuracy of 86%, precision of 95% and recall of 88%. Four out of five analyzed proteins showed a noticeable difference between the results of docking ligands labeled as promising and those from the bottom of the list returned by the neural network.

Housing price prediction

---

Based on tabular data, including information such as for example, residential size, number of bedrooms, number of garage spaces, number of balconies, location, condition of the apartment, etc., we created a machine learning model that predicts the value of the apartment/home. The model performed very well predicting the value of the property with an accuracy of 2% of the value.

Fetal Biometry Estimation

---

We have developed a deep learning model that can predict measurements of the femur, abdomen, and head based on ultrasound sonography (USG) scans of fetuses. Our model holds significant importance in healthcare for several reasons:

Efficiency and Time-saving: The automation of fetal biometry through our model can significantly reduce the time required for manual measurements, enabling healthcare providers to cater to more patients or allocate more time to other tasks.

Consistency: Our automated model can provide consistent measurements, effectively eliminating the variability introduced by different technicians performing manual measurements.

Accuracy: Our model leverages advanced machine learning algorithms, demonstrating the potential to achieve high levels of accuracy, which could reduce measurement errors.

Early Detection: With our automated biometry model, fetal growth can be assessed more frequently. This could allow for earlier detection of growth abnormalities or complications.